ahmetbengoz/horizon-europe-revealed-importance: DSS revision release – outcome-conditioned proxy-profile framework for Horizon Europe

Horizon Europe Funding-Instrument Differentiation — v3.0.0 This release establishes a new research identity for the repository: Measuring funding-instrument differentiation through observable project portfolios: Evidence from Horizon Europe The study evaluates how strongly and consistently six Horizon Europe instrument families—MSCA, ERC, RIA, CSA, IA and EIC—differ in the observable characteristics of their funded-project portfolios. Scientific redesign Replaces the former proxy-importance and learned-weight interpretation. Defines funding-instrument differentiation as a descriptive evaluation diagnostic. Introduces a complete variable-role and data-leakage audit. Excludes funding schemes, call and topic codes, legal basis, identifiers, titles and status from model predictors. Uses project scale, collaboration, organisation mix, thematic orientation and text-structure features. Makes no claims about evaluator preferences, official weights, selection quality or causal programme effects. Analysis Official CORDIS Horizon Europe snapshot served on 6 August 2026. 23,451 unique project records. Primary cohort: 17,866 projects starting in 2021–2025. Forward-time validation: training on 2021–2024 and testing on 4,453 projects starting in 2025. Robust univariate effects and PERMANOVA. Logistic regression, random forest, Extra Trees and histogram gradient boosting benchmarks. Calibration, grouped permutation importance, ablation and robustness analyses. Provisional 2026–2027 cohort retained only as a sensitivity test. Key frozen results PERMANOVA pseudo-F = 699.81, R² = 0.164, p = 0.001. 2025 holdout macro-F1 = 0.851. 2025 balanced accuracy = 0.849. Collaborative-family subset macro-F1 = 0.825. Non-financial specification macro-F1 = 0.733. Provisional 2026–2027 holdout macro-F1 = 0.774. Reproducibility This release includes: end-to-end analysis scripts; deterministic repair of malformed CORDIS rows; variable-role and leakage diagnostics; machine-readable CSV results; 600 dpi publication figures; release and independent Zenodo-deposition notes. Raw and processed row-level CORDIS data are not redistributed. The scripts download the official third-party source data. Previous repository identity The superseded proxy-profile workflow and outputs are preserved under: archive/v2_proxy_profiles/ They are retained for provenance only and must not be used to support the new article's claims. Frozen commit 5bbc42c811261cbfe172077025aec2408d22be71